Reduce the memory footprint of dsymutil. (NFC)

This (partially) fixes a regression introduced by
https://reviews.llvm.org/D43945 / r327399, which parallelized
DwarfLinker. This patch avoids parsing and allocating the memory for
all input DIEs up front and instead only allocates them in the
concurrent loop in the AnalyzeLambda. At the end of the loop the
memory from the LinkContext is cleared again.

This reduces the peak memory needed to link the debug info of a
non-modular build of the Swift compiler by >3GB.

rdar://problem/43444464

Differential Revision: https://reviews.llvm.org/D51078

llvm-svn: 340650